<?xml version="1.0" encoding="UTF-8"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom" xmlns:dc="http://purl.org/dc/elements/1.1/">
  <channel>
    <title>DEV Community: FreshStore</title>
    <description>The latest articles on DEV Community by FreshStore (freshstore).</description>
    <link>https://dev.to/freshstore</link>
    <image>
      <url>https://media2.dev.to/dynamic/image/width=90,height=90,fit=cover,gravity=auto,format=auto/https:%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Forganization%2Fprofile_image%2F16228%2F0e7963fe-16cd-43da-ae3d-a136f24a9f45.png</url>
      <title>DEV Community: FreshStore</title>
      <link>https://dev.to/freshstore</link>
    </image>
    <atom:link rel="self" type="application/rss+xml" href="https://dev.to/feed/freshstore"/>
    <language>en</language>
    <item>
      <title>How I run a 16-year-old SaaS with AI agents</title>
      <dc:creator>Carey</dc:creator>
      <pubDate>Wed, 18 Nov 2026 10:52:07 +0000</pubDate>
      <link>https://dev.to/freshstore/how-i-run-a-16-year-old-saas-with-ai-agents-oli</link>
      <guid>https://dev.to/freshstore/how-i-run-a-16-year-old-saas-with-ai-agents-oli</guid>
      <description>&lt;p&gt;Most people think AI in a SaaS means a chatbot in the corner of the pricing page.&lt;/p&gt;

&lt;p&gt;That's the least useful place for it.&lt;/p&gt;

&lt;p&gt;I've run &lt;a href="https://www.freshstore.com/" rel="noopener noreferrer"&gt;FreshStore&lt;/a&gt;, an affiliate site builder, since 2010. It's a small team. Over the last year the biggest change has been in how the business runs behind the product. AI agents now watch the servers, triage support, work the bug list and write the first draft of most of the boring things.&lt;/p&gt;

&lt;h2&gt;
  
  
  Cheap checks first, smart agents second
&lt;/h2&gt;

&lt;p&gt;My first attempt was the obvious one. Point an agent at the logs every hour and ask it if anything looked wrong.&lt;/p&gt;

&lt;p&gt;It was slow, it was expensive, and it got bored. Most hours nothing is wrong, so most runs were an agent reading healthy logs and saying so.&lt;/p&gt;

&lt;p&gt;What works is two layers. The first layer is plain scripts. They run all the time, they cost almost nothing, and they only answer simple questions. Is the site up? Did the error count jump? Did a payment fail? Did a queue stop moving?&lt;/p&gt;

&lt;p&gt;The agent is the second layer. It only wakes up when a script trips, or once a day for a proper look. Then it gets the context, digs in, and decides whether it's real.&lt;/p&gt;

&lt;h2&gt;
  
  
  Findings go on a list, not in a chat
&lt;/h2&gt;

&lt;p&gt;The second mistake was alerts. Every agent wanted to tell me something, and they all wanted to tell me now.&lt;/p&gt;

&lt;p&gt;Now an agent files a finding. It's a task with a title, a priority, the evidence, and a stable key. If the same problem comes back, it reopens the same task rather than creating a new one. Only real emergencies, a site down or money being lost, ping my phone.&lt;/p&gt;

&lt;p&gt;That one change took me from dozens of messages a day to a list I can read with a coffee.&lt;/p&gt;

&lt;h2&gt;
  
  
  Never ignore an error
&lt;/h2&gt;

&lt;p&gt;This is the rule I'm most stubborn about.&lt;/p&gt;

&lt;p&gt;It's tempting to build a list of "known" errors and tell the agents to skip them. I don't. A recurring error is a bug nobody has fixed yet. The agents can decide when something is worth waking up for. They never get to decide that something isn't worth fixing.&lt;/p&gt;

&lt;h2&gt;
  
  
  A fixer that works the list
&lt;/h2&gt;

&lt;p&gt;Once findings are on a list, something has to clear them.&lt;/p&gt;

&lt;p&gt;A separate agent works through the open findings. It rechecks each one against the live system first, because a surprising number fix themselves or turn out to be something else. If the fix is safe and small, it makes it, tests it, ships it and writes down what it did. If it needs a decision about money, members or how the product should work, it stops and leaves me a short note with a recommendation.&lt;/p&gt;

&lt;p&gt;That split matters more than any prompt. The agent does the work and I keep the decisions.&lt;/p&gt;

&lt;h2&gt;
  
  
  Guardrails are code, not instructions
&lt;/h2&gt;

&lt;p&gt;Telling an agent "never do X" works most of the time. Most of the time isn't good enough when X is refunding a customer.&lt;/p&gt;

&lt;p&gt;So the hard rules are enforced in code. Before an agent's action runs, a small check looks at it and blocks anything on the wrong side of a line. Billing changes, sending a customer email, deleting things. The agent can draft all of it. It can't do it without me.&lt;/p&gt;

&lt;p&gt;The nice side effect is that the rules are written down in one place. When I change my mind, I change one file, and every agent follows it from the next run.&lt;/p&gt;

&lt;h2&gt;
  
  
  Support: drafts, never sends
&lt;/h2&gt;

&lt;p&gt;Support was the area I was most nervous about, and it's turned out to be one of the best.&lt;/p&gt;

&lt;p&gt;An agent reads each new ticket, looks up the member's account, checks the logs for what actually happened, and writes a draft reply. A person reads it and sends it. The draft is usually right. When it isn't, it's still faster to fix a draft that has already done the digging than to start from a blank reply.&lt;/p&gt;

&lt;h2&gt;
  
  
  Everything leaves a trail
&lt;/h2&gt;

&lt;p&gt;Every agent run writes down what it looked at, what it changed and why. Every shipped change says whether members can see it, and a draft line for the public changelog falls out of that for free.&lt;/p&gt;

&lt;p&gt;At the end of the day I read short briefs, not logs. If something looks off, the trail is there.&lt;/p&gt;

&lt;h2&gt;
  
  
  What didn't work
&lt;/h2&gt;

&lt;p&gt;A few things I'd skip next time:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;One giant agent that does everything.&lt;/strong&gt; It loses the plot. Small agents with one job each are easier to trust and easier to fix.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Long-running sessions.&lt;/strong&gt; Context goes stale. A fresh run each time, with notes written to disk, beats one agent that's been awake all week.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Trusting "done".&lt;/strong&gt; An agent saying a fix is live isn't proof. Every deploy now ends with a real check against the live site.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  Where this is going
&lt;/h2&gt;

&lt;p&gt;This moved the work rather than removing it. I spend less time reading logs and chasing bugs, and more time on the decisions only I can make.&lt;/p&gt;

&lt;p&gt;If I were starting again, I wouldn't begin with a chatbot. I'd begin with the thing I check every morning, and teach something to check it for me.&lt;/p&gt;

</description>
      <category>ai</category>
      <category>saas</category>
      <category>devops</category>
      <category>productivity</category>
    </item>
  </channel>
</rss>
